On 64-bit x86, emitting Assembler::pusha and ::popa cost roughly 1.2M instructions during startup. These assembler routines only emit non-relocating instructions and are invariant with any environment settings (UseAVX etc..), so the instruction stream could be pre-calculated and streamlined considerably.

PoC experiment: http://cr.openjdk.java.net/~redestad/scratch/pusha_popa.00/

Before:
       111,118,300 instructions # 0.80 insns per cycle ( +- 0.07% )
        21,941,371 branches # 414.614 M/sec ( +- 0.07% )
           766,596 branch-misses # 3.49% of all branches ( +- 0.15% )

After:
       110,039,451 instructions # 0.81 insns per cycle ( +- 0.07% )
        21,792,132 branches # 419.041 M/sec ( +- 0.07% )
           761,485 branch-misses # 3.49% of all branches ( +- 0.14% )